A Kalman Filter Approach to Correct Surface Forecast Bias

Despite improvements in numerical weather prediction, model biases are still unavoidable due to imperfect model physics, initialization, and boundary conditions. Recent studies (e.g., Hart et al. 2004) suggest that statistical post-processing of model forecasts may help to reduce model biases. We have chosen to use the linear Kalman filter (KF) approach to correct the surface wind and temperature forecast biases from the Eta model for approximately 2400 stations in the Western United States, following the approach of Roeger et al. (2003). The KF is essentially a predictor-corrector type estimator that minimizes the estimated error covariance, using both observations and previous estimates. Advantages of the KF include: i) conceptual simplicity and ii) the system “learning” from its past. The observational database used in conjunction with the KF is the MesoWest, a high-density surface observational network with over 3000 stations in the Western United States maintained by the Cooperative Institute for Regional Prediction at University of Utah. By using MesoWest, point-specific forecasts can be made in many regions where traditional Model Output Statistics (MOS) are unavailable, especially upper-elevation mountain and mountain valley locations that are far removed from airports.

Of note, the KF-corrected forecasts may be particularly useful in areas of complex terrain, such as in the Western United States, where models are unable to accurately represent the underlying orography. In addition, because of poor model resolution and physics, numerical models perform poorly in some weather phenomena in areas of complex terrain, such as persistent cold pool events in valleys during winter. Due to the retention of the memory of the bias errors, the KF-corrected surface forecasts do well in persistent situations such as the one just described when full physics numerical models do not perform as well.

For this study, we verified the 48-h KF-corrected Eta surface temperature and wind forecasts for the 2003/2004 winter season (December 2003 to February 2004) and the 2004 summer season (June to August 2004) for the 0000 and 1200 UTC cycles against MesoWest observations. In addition, we also compared the performance of the KF-corrected Eta forecasts to MOS from the Eta, NGM and GFS models, as well as a persistence-bias corrected Eta forecasts and the original uncorrected Eta forecasts at selected 150 stations. Development is underway to provide over 2400 KF-corrected point forecasts in the Western United States to the Weather Forecasting Offices, as part of a collaborative project with the National Weather Service.